how to filter invalid pairs?

[tengteng95]

I found the segmentation result of person image sometimes is annoying. For example, for the upper clothes, there may output several different labels(such as upper_clothes, dress, skirt etc.) As a result, it's hard to decide to take which part as the real upper clothes region.

Could you please show the code for processing this kind of cases? Thanks a lot.

[xthan]

I did not process these cases. I only use upper_clothes in the segmentation map as the clothing region as written in:
https://github.com/xthan/VITON/blob/master/utils.py#L58

[tengteng95]

Thanks for your reply. I'm afraid that only using 'upper_clothes' will result in bad segmentation results and be bad for Shape Context.
By the way, do you only filter those pairs with no face detected? As is shown in your paper, there are around 3k images filtered.

[xthan]

I filter the images with bad upper_clothes segmentation results by looking at the area of the segment - if the upper_clothes region only occupies a small portion of the image, I regard this as a bad detection and discard it. There is no face detection used.

[tengteng95]

Thanks!